GPT-5.4 nears first FrontierMath solve

If this claim survives verification, it matters more than another benchmark flex because FrontierMath open problems are supposed to look like real mathematical research, not polished test prep. The bigger story is that model progress is starting to outrun the community’s ability to validate whether an AI result counts as genuine new mathematics.

• –Epoch describes FrontierMath open problems as unsolved questions that resisted serious attempts by professional mathematicians and would meaningfully advance human mathematical knowledge if solved
• –The Reddit thread quotes an X claim that the result came from a single GPT-5.4 Pro run and was later refined into Lean with a higher-compute GPT-5.4 setting
• –Epoch’s own changelog is the key caution flag: it recently removed one problem after deciding an AI-generated solution did not meet the benchmark’s bar for a publishable result
• –If confirmed, this would push math evaluation beyond olympiad-style scoring and into “can the model generate original research artifacts” territory